perf/x86: Implement IBS pmu control ops
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / arch / x86 / kernel / cpu / perf_event_amd_ibs.c
index 3b8a2d30d14e8ebeb2c58406212fbbd3c79ba935..40a6d9d5dd237599780cb47cef602ea99c56efcc 100644 (file)
@@ -16,36 +16,298 @@ static u32 ibs_caps;
 
 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD)
 
-static struct pmu perf_ibs;
+#include <linux/kprobes.h>
+#include <linux/hardirq.h>
+
+#include <asm/nmi.h>
+
+#define IBS_FETCH_CONFIG_MASK  (IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
+#define IBS_OP_CONFIG_MASK     IBS_OP_MAX_CNT
+
+enum ibs_states {
+       IBS_ENABLED     = 0,
+       IBS_STARTED     = 1,
+       IBS_STOPPING    = 2,
+
+       IBS_MAX_STATES,
+};
+
+struct cpu_perf_ibs {
+       struct perf_event       *event;
+       unsigned long           state[BITS_TO_LONGS(IBS_MAX_STATES)];
+};
+
+struct perf_ibs {
+       struct pmu      pmu;
+       unsigned int    msr;
+       u64             config_mask;
+       u64             cnt_mask;
+       u64             enable_mask;
+       u64             valid_mask;
+       unsigned long   offset_mask[1];
+       int             offset_max;
+       struct cpu_perf_ibs __percpu *pcpu;
+};
+
+struct perf_ibs_data {
+       u32             size;
+       union {
+               u32     data[0];        /* data buffer starts here */
+               u32     caps;
+       };
+       u64             regs[MSR_AMD64_IBS_REG_COUNT_MAX];
+};
+
+static struct perf_ibs perf_ibs_fetch;
+static struct perf_ibs perf_ibs_op;
+
+static struct perf_ibs *get_ibs_pmu(int type)
+{
+       if (perf_ibs_fetch.pmu.type == type)
+               return &perf_ibs_fetch;
+       if (perf_ibs_op.pmu.type == type)
+               return &perf_ibs_op;
+       return NULL;
+}
 
 static int perf_ibs_init(struct perf_event *event)
 {
-       if (perf_ibs.type != event->attr.type)
+       struct hw_perf_event *hwc = &event->hw;
+       struct perf_ibs *perf_ibs;
+       u64 max_cnt, config;
+
+       perf_ibs = get_ibs_pmu(event->attr.type);
+       if (!perf_ibs)
                return -ENOENT;
+
+       config = event->attr.config;
+       if (config & ~perf_ibs->config_mask)
+               return -EINVAL;
+
+       if (hwc->sample_period) {
+               if (config & perf_ibs->cnt_mask)
+                       /* raw max_cnt may not be set */
+                       return -EINVAL;
+               if (hwc->sample_period & 0x0f)
+                       /* lower 4 bits can not be set in ibs max cnt */
+                       return -EINVAL;
+               max_cnt = hwc->sample_period >> 4;
+               if (max_cnt & ~perf_ibs->cnt_mask)
+                       /* out of range */
+                       return -EINVAL;
+               config |= max_cnt;
+       } else {
+               max_cnt = config & perf_ibs->cnt_mask;
+               event->attr.sample_period = max_cnt << 4;
+               hwc->sample_period = event->attr.sample_period;
+       }
+
+       if (!max_cnt)
+               return -EINVAL;
+
+       hwc->config_base = perf_ibs->msr;
+       hwc->config = config;
+
        return 0;
 }
 
+static void perf_ibs_start(struct perf_event *event, int flags)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
+       struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
+
+       if (test_and_set_bit(IBS_STARTED, pcpu->state))
+               return;
+
+       wrmsrl(hwc->config_base, hwc->config | perf_ibs->enable_mask);
+}
+
+static void perf_ibs_stop(struct perf_event *event, int flags)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
+       struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
+       u64 val;
+
+       if (!test_and_clear_bit(IBS_STARTED, pcpu->state))
+               return;
+
+       set_bit(IBS_STOPPING, pcpu->state);
+
+       rdmsrl(hwc->config_base, val);
+       val &= ~perf_ibs->enable_mask;
+       wrmsrl(hwc->config_base, val);
+}
+
 static int perf_ibs_add(struct perf_event *event, int flags)
 {
+       struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
+       struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
+
+       if (test_and_set_bit(IBS_ENABLED, pcpu->state))
+               return -ENOSPC;
+
+       pcpu->event = event;
+
+       if (flags & PERF_EF_START)
+               perf_ibs_start(event, PERF_EF_RELOAD);
+
        return 0;
 }
 
 static void perf_ibs_del(struct perf_event *event, int flags)
 {
+       struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
+       struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
+
+       if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
+               return;
+
+       perf_ibs_stop(event, 0);
+
+       pcpu->event = NULL;
 }
 
-static struct pmu perf_ibs = {
-       .event_init= perf_ibs_init,
-       .add= perf_ibs_add,
-       .del= perf_ibs_del,
+static void perf_ibs_read(struct perf_event *event) { }
+
+static struct perf_ibs perf_ibs_fetch = {
+       .pmu = {
+               .task_ctx_nr    = perf_invalid_context,
+
+               .event_init     = perf_ibs_init,
+               .add            = perf_ibs_add,
+               .del            = perf_ibs_del,
+               .start          = perf_ibs_start,
+               .stop           = perf_ibs_stop,
+               .read           = perf_ibs_read,
+       },
+       .msr                    = MSR_AMD64_IBSFETCHCTL,
+       .config_mask            = IBS_FETCH_CONFIG_MASK,
+       .cnt_mask               = IBS_FETCH_MAX_CNT,
+       .enable_mask            = IBS_FETCH_ENABLE,
+       .valid_mask             = IBS_FETCH_VAL,
+       .offset_mask            = { MSR_AMD64_IBSFETCH_REG_MASK },
+       .offset_max             = MSR_AMD64_IBSFETCH_REG_COUNT,
 };
 
+static struct perf_ibs perf_ibs_op = {
+       .pmu = {
+               .task_ctx_nr    = perf_invalid_context,
+
+               .event_init     = perf_ibs_init,
+               .add            = perf_ibs_add,
+               .del            = perf_ibs_del,
+               .start          = perf_ibs_start,
+               .stop           = perf_ibs_stop,
+               .read           = perf_ibs_read,
+       },
+       .msr                    = MSR_AMD64_IBSOPCTL,
+       .config_mask            = IBS_OP_CONFIG_MASK,
+       .cnt_mask               = IBS_OP_MAX_CNT,
+       .enable_mask            = IBS_OP_ENABLE,
+       .valid_mask             = IBS_OP_VAL,
+       .offset_mask            = { MSR_AMD64_IBSOP_REG_MASK },
+       .offset_max             = MSR_AMD64_IBSOP_REG_COUNT,
+};
+
+static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
+{
+       struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
+       struct perf_event *event = pcpu->event;
+       struct hw_perf_event *hwc = &event->hw;
+       struct perf_sample_data data;
+       struct perf_raw_record raw;
+       struct pt_regs regs;
+       struct perf_ibs_data ibs_data;
+       int offset, size;
+       unsigned int msr;
+       u64 *buf;
+
+       if (!test_bit(IBS_STARTED, pcpu->state)) {
+               /* Catch spurious interrupts after stopping IBS: */
+               if (!test_and_clear_bit(IBS_STOPPING, pcpu->state))
+                       return 0;
+               rdmsrl(perf_ibs->msr, *ibs_data.regs);
+               return (*ibs_data.regs & perf_ibs->valid_mask) ? 1 : 0;
+       }
+
+       msr = hwc->config_base;
+       buf = ibs_data.regs;
+       rdmsrl(msr, *buf);
+       if (!(*buf++ & perf_ibs->valid_mask))
+               return 0;
+
+       perf_sample_data_init(&data, 0);
+       if (event->attr.sample_type & PERF_SAMPLE_RAW) {
+               ibs_data.caps = ibs_caps;
+               size = 1;
+               offset = 1;
+               do {
+                   rdmsrl(msr + offset, *buf++);
+                   size++;
+                   offset = find_next_bit(perf_ibs->offset_mask,
+                                          perf_ibs->offset_max,
+                                          offset + 1);
+               } while (offset < perf_ibs->offset_max);
+               raw.size = sizeof(u32) + sizeof(u64) * size;
+               raw.data = ibs_data.data;
+               data.raw = &raw;
+       }
+
+       regs = *iregs; /* XXX: update ip from ibs sample */
+
+       if (perf_event_overflow(event, &data, &regs))
+               ; /* stop */
+       else
+               /* reenable */
+               wrmsrl(hwc->config_base, hwc->config | perf_ibs->enable_mask);
+
+       return 1;
+}
+
+static int __kprobes
+perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
+{
+       int handled = 0;
+
+       handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
+       handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
+
+       if (handled)
+               inc_irq_stat(apic_perf_irqs);
+
+       return handled;
+}
+
+static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
+{
+       struct cpu_perf_ibs __percpu *pcpu;
+       int ret;
+
+       pcpu = alloc_percpu(struct cpu_perf_ibs);
+       if (!pcpu)
+               return -ENOMEM;
+
+       perf_ibs->pcpu = pcpu;
+
+       ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
+       if (ret) {
+               perf_ibs->pcpu = NULL;
+               free_percpu(pcpu);
+       }
+
+       return ret;
+}
+
 static __init int perf_event_ibs_init(void)
 {
        if (!ibs_caps)
                return -ENODEV; /* ibs not supported by the cpu */
 
-       perf_pmu_register(&perf_ibs, "ibs", -1);
+       perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
+       perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
+       register_nmi_handler(NMI_LOCAL, &perf_ibs_nmi_handler, 0, "perf_ibs");
        printk(KERN_INFO "perf: AMD IBS detected (0x%08x)\n", ibs_caps);
 
        return 0;